WebWorld: Evolving Ecology & Web Simulation

• WebWorld is a multifaceted research area comprising evolving ecology models, large-scale web simulators, and synthetic web environments with distinct methodologies.
• It leverages directed, weighted graph analysis and link-shuffle controls to assess complexity surpluses in foodweb-like ecological networks.
• WebWorld research employs autoregressive LLMs and explicit state-transition frameworks to generate synthetic trajectories and enhance web-agent training.

WebWorld is used for multiple technically distinct constructs in the research literature. In artificial-life and network-complexity studies, Webworld denotes an evolving ecology model whose foodweb-like interaction structure is analyzed as a directed, weighted graph; in that setting it became notable because it did not show the complexity-surplus pattern that had been hypothesized as a signature of evolutionary dynamics (Standish, 2010, Standish, 2014). In recent web-agent research, WebWorld denotes a large-scale open-web world model that predicts browser state transitions from interaction history and is trained on 1,059,348 trajectories to support synthetic trajectory generation, inference-time search, and downstream agent training (Xiao et al., 16 Feb 2026). Related work broadens the notion further toward controllable finite-state websites and code-backed web-native worlds (Wu et al., 15 Feb 2026, Feng et al., 29 Dec 2025).

1. Terminological scope

The term spans at least three research contexts.

Usage	Modeled object	Characteristic role
Webworld in artificial life	evolving ecology / foodweb	source of directed, weighted ecological networks
WebWorld in web-agent modeling	browser state transitions	open-web simulator trained at scale
broader WebWorld vision	synthetic or code-backed web environments	controllable worlds for training, planning, or persistent interaction

In the first sense, Standish studies Webworld alongside Tierra and EcoLab as an artificial-life source of ecological networks (Standish, 2010). In the second, WebWorld is introduced as an open-web simulator trained at scale, with 8B, 14B, and 32B variants, 30+ turn simulation, multi-format support, and a 1.06M-trajectory corpus (Xiao et al., 16 Feb 2026). In the third, AutoWebWorld frames a broader “WebWorld” idea in which web environments are constructed with explicit states, actions, and transition rules, while the Web World Model framework places code-defined state and invariants inside ordinary web stacks (Wu et al., 15 Feb 2026, Feng et al., 29 Dec 2025). This suggests a shift from treating WebWorld as a generator of analyzable networks to treating it as a programmable substrate for agent environments.

2. Webworld as an evolving ecology model

In Standish’s 2010 study, Webworld is described as an evolving ecology model originally introduced by Caldarelli et al. and modified by Drossel et al. It is said to be “more realistic” than EcoLab in the sense that it explicitly tracks biomass resources. Like EcoLab, it has an interaction matrix, called the functional response in Webworld, which can be interpreted as a foodweb; for complexity analysis this matrix is converted into a directed, weighted graph (Standish, 2010).

The methodological role of Webworld is comparative rather than foundational. Standish uses it to test the hypothesis that the complexity surplus seen in several real food webs reflects information accumulated by evolution. For each sampled Webworld foodweb, the analysis generates 100 link-shuffled control versions, computes the average and standard deviation of $\ln {\cal C}$ rather than ${\cal C}$, and compares the evolved network with the geometric mean of its shuffled controls through

$$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$

The reported result is that Webworld food webs show very little information difference from their shuffled counterparts; none of the artificial-life models studied produce a surplus substantial enough to exceed about 10% of total complexity, and the observed differences are often only a few bits (Standish, 2010).

The later 2014 paper sharpens the interpretation. After correcting a bug in an earlier analysis, EcoLab and Tierra are reported to show significant complexity surpluses, “though not for WebWorld, another evolutionary ecology model similar to EcoLab.” WebWorld therefore becomes an explicit contrast case: it is evolutionary, yet it does not display a significant surplus (Standish, 2014). A common but insufficient inference—that complexity surplus is a generic marker of evolutionary dynamics—is thereby weakened.

3. Information-theoretic complexity and the surplus criterion

The complexity framework used in the Webworld studies is information-theoretic. Standish defines

$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$

where $\omega(\ell,x)$ is the number of descriptions of length $\ell$ that map to the same object $x$, and $N$ is the alphabet size of the prefix-free representation language (Standish, 2010). In the 2014 formulation this is also written as

$${\cal C}(x)=-\log P(x)=\lim_{\ell\rightarrow\infty}(\ell\log N-\log\omega(\ell,x)).$$

For graphs encoded in a binary alphabet, the concrete representation specifies a prefix encoding of the number of nodes $n$, an encoding of the number of links ${\cal C}$0, and then a rank-index encoding of the link list among all possible link lists with ${\cal C}$1 links. This yields

${\cal C}$2

The observer or classifier identifies topologically identical graphs under relabeling, so the measure counts equivalent descriptions rather than merely the raw size of an adjacency structure (Standish, 2014).

Weighted networks require an extension. The 2010 paper writes

${\cal C}$3

with the practical computation performed by summing the complexities of partial networks obtained by including links in weight order (Standish, 2010). The 2014 paper describes the same idea procedurally: form graphs consisting of links with weight at least ${\cal C}$4, compute graph complexity for each thresholded graph, and then take a weighted sum over the distinct weights (Standish, 2014).

Within this framework, the decisive question is not the absolute complexity of a network but whether it exceeds that of a shuffled Erdős–Rényi-style counterpart preserving the original link-weight distribution. Webworld matters because it fails this test despite being an evolutionary ecology model. That negative result is central to the later claim that complexity surplus is more fundamental than “being an evolutionary system” (Standish, 2014).

4. World models in web navigation

A separate research line reintroduces WebWorld in the context of web agents. Here the underlying problem is long-horizon web interaction under partial observability, sparse rewards, and costly real-world rollout. The 2024 World-model-Augmented web agent paper argues that current LLM-based web agents lack a world model: preliminary analyses on WebArena show that GPT-4o, GPT-4o-mini, GPT-4-Turbo, and Claude-3.5-Sonnet achieve only about 54.75% average next-state classification accuracy, while providing explicit candidate outcomes improves action selection by as much as 38 percentage points (Chae et al., 2024).

The proposed remedy is model-based action selection. A frozen policy model ${\cal C}$5 proposes candidate actions, a learned world model ${\cal C}$6 predicts their consequences, and a value function ${\cal C}$7 scores the imagined next states: ${\cal C}$8 A key technical device is the transition-focused observation abstraction, which uses the Hungarian algorithm to summarize only the important differences between ${\cal C}$9 and $$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$0 rather than predicting full next-page text. On WebArena, this world-model augmentation raises GPT-4o from 13.1% to 16.6% success rate, while remaining about 6.8$$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$1 cheaper and 5.3$$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$2 faster than a tree-search agent (Chae et al., 2024).

OpenWebVoyager addresses a different bottleneck: iterative self-improvement in real, multimodal web settings. It bootstraps an open-source multimodal backbone with imitation learning from WebVoyager-4o trajectories, then repeats an exploration–feedback–optimization cycle on real websites. The reported WebVoyager overall success rate rises from 19.9% after imitation learning to 25.8% by iter-3, while Mind2Web cross-task rises from 6.3% to 19.6% (He et al., 2024). This establishes that real-web improvement loops can work without an explicit simulator, but they remain tied to expensive live interaction.

WebEvolver adds a coevolving world model to such self-improvement loops. It formalizes web interaction as a POMDP, trains a World Model LLM $$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$3 to predict next observations, uses it both as a virtual web server for synthetic rollouts and as an imagination engine for inference-time planning, and selects actions by

$$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$4

The reported overall success rates are 42.49 on WebVoyager and 22.64 on Mind2Web-Live, compared with 38.23 and 16.98 for the final self-improve baseline, while World Model Look-Ahead with depth $$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$5 reaches 51.37 on WebVoyager and 24.53 on Mind2Web-Live (Fang et al., 23 Apr 2025). The same paper also reports short-horizon degradation beyond depth 2, indicating that planning quality remains tightly coupled to world-model fidelity.

5. WebWorld as a large-scale open-web simulator

The 2026 WebWorld series makes the world model itself the central object. WebWorld is introduced as the first open-web simulator trained at scale, designed to predict how a browser page changes after an action so that agents can train, search, and synthesize trajectories without constant live-web interaction. The model is autoregressive: $$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$6 with training objective

$$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$7

It is instantiated with causal LLMs at 8B, 14B, and 32B parameters (Xiao et al., 16 Feb 2026).

The data pipeline is hierarchical and explicitly pretraining-aligned. URLs are sourced from FineWeb and filtered CCI 3.0; randomized crawling yields 293,511 trajectories, autonomous exploration 38,356, task-oriented execution 94,001, open-source data 37,568, multi-format data 47,855, and interaction data 547,758, for a total of 1,059,348 trajectories (Xiao et al., 16 Feb 2026).

Source	Trajectories	Role
Randomized Crawling	293,511	short executable exploration
Autonomous Exploration	38,356	long-horizon self-proposed goals
Task-Oriented Execution	94,001	successful goal-directed traces
Open Source	37,568	external trajectory enrichment
Multi-format	47,855	representation robustness
Interaction data	547,758	corpus expansion
Total	1,059,348	training set

The training scheme has two stages. First, transition modeling learns $$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$8 over the full corpus. Second, a CoT activation stage uses 1K synthesized CoT samples to teach $$\Delta = {\cal C} - e^{\langle \ln {\cal C}_\mathrm{ER}\rangle}.$$9. The paper argues that strong dynamics learned from the large dataset allow this small reasoning set to activate explicit chain-of-thought effectively; it further states that 1,000 CoT samples suffice to outperform direct reasoning tuning on Qwen3-8B with 10$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$0 more data (Xiao et al., 16 Feb 2026).

Intrinsic evaluation is organized around WebWorld-Bench, which uses two GPT-4o-based metrics—Factuality Score and Web Turing Score—across nine dimensions: Long-Horizon Consistency, Base Semantics, Fine-Grained Sensitivity, Multi-tab / Multi-page, Format Robustness: XML, Format Robustness: HTML, Format Robustness: Markdown, Format Robustness: Playwright, and Web2NAL. On this benchmark, WebWorld-32B obtains an average Factuality Score of 71.0%, compared with 70.3% for Gemini-3-Pro and 71.3% for Claude-Opus-4.1, with 77.0% on long-horizon factuality (Xiao et al., 16 Feb 2026).

Extrinsically, WebWorld synthesizes 8,000 diverse trajectories via Abstract-and-Instantiate and fine-tunes Qwen3 agents. For Qwen3-14B, WebArena rises from 15.1% to 24.3%, an improvement of +9.2%, which the paper summarizes as performance comparable to GPT-4o. For Qwen3-8B, WebArena rises from 9.8% to 20.7%, and MiniWob++ from 49.4% to 59.3%. The same paper reports that WebWorld can outperform GPT-5 as a world model in lookahead search, but also states that search yields only modest gains overall and that world models are more valuable as training-data generators than as a planning crutch. Its limitations include sycophancy / optimism bias, weak high-quality content generation, privacy risk, text-centric visual limitations, and the observation that inference-time search is not a panacea (Xiao et al., 16 Feb 2026).

6. Synthetic and verifiable web environments

A parallel direction replaces learned web dynamics with explicit symbolic construction. AutoWebWorld models a web environment as a deterministic system

$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$1

with semantic state

$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$2

where $${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$3 is a page id and $${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$4 is a page signature containing structured task-relevant variables such as query text, filters, pagination index, selected items, cart contents, and form field values (Wu et al., 15 Feb 2026).

Transitions are verified internally. Invalid actions satisfy $${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$5 and typically become no-ops, $${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$6; valid actions apply deterministic effects, either as in-page updates

$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$7

or as cross-page transitions

$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$8

Goal states are explicit and can be written, for example, as

$${\cal C}(x)=\lim_{\ell\rightarrow\infty}\ell\log N-\log\omega(\ell,x),$$9

Because state, action legality, and transition rules are explicit, step correctness and task success are programmatically decidable rather than inferred from screenshots or external judges (Wu et al., 15 Feb 2026).

The generation pipeline uses a multi-agent FSM authoring loop, a coding-agent translation stage, breadth-first search over the semantic state graph with deduplication key

$\omega(\ell,x)$0

expands each high-level action into a gui_procedure, and replays candidates in Playwright for execution-based filtering. The reported scale is 29 synthetic web environments, 875 website pages, 11,663 verified trajectories, average trajectory length about 21.9 steps, and \$0.04 per trajectory. After sampling one representative trajectory from parallel variants, the GRPO training set contains 1,215 distinct trajectories, 12,585 interaction steps, and about 16k total training steps. On WebVoyager, Ours-7B reaches 27.42% overall, above UI-TARS-1.5-7B at 26.51%, and Online-Mind2Web shows monotonic gains as synthetic data increases from 1.22% at 8 samples to 14.02% at 16,253 samples (Wu et al., 15 Feb 2026).

InfiniteWeb addresses the same scarcity problem from an environment-synthesis rather than FSM-first perspective. It automatically generates functional multi-page websites, tasks, shared backend/data specifications, page architectures, and evaluators. Its task-centric test-driven development (TCTDD) iterates through test generation, code generation, execution, failure inspection, and repair, up to a maximum of 8 iterations. A major contribution is evaluator generation for dense rewards, expressed as

$\omega(\ell,x)$1

with explicit checkpoint examples such as subscription attempt validity, storage consistency, and on-site confirmation (Zhang et al., 7 Jan 2026).

Dense rewards substantially change the effective training signal: the paper reports learning from 767 tasks versus 174 tasks with binary reward, a 4.4$\omega(\ell,x)$2 increase in discriminative tasks. On WebGen-Bench, InfiniteWeb achieves 85.6 overall functional score, above Bolt.diy (67.1), Claude-Code (75.8), and Codex (80.8). Training UI-TARS-1.5-7B on 600 generated tasks with GRPO improves OSWorld from 24.5% to 31.4% and yields +5.7% on Online-Mind2Web (Zhang et al., 7 Jan 2026). Together, AutoWebWorld and InfiniteWeb show that a WebWorld need not be mined from the live internet; it can be synthesized, verified, and scaled by construction.

7. Web-native world models as software architecture

The Web World Model framework identifies a middle ground between fixed conventional web applications and fully generative world models. Its central decomposition is

$\omega(\ell,x)$3

where the Physics layer $\omega(\ell,x)$4 is deterministic and code-defined, and the Imagination layer $\omega(\ell,x)$5 is stochastic and model-defined. State transitions are ordered: $\omega(\ell,x)$6 The model is therefore not permitted to invent the world arbitrarily; it must condition on a code-validated symbolic state that preserves invariants such as inventories, coordinates, resources, connectivity, or legal actions (Feng et al., 29 Dec 2025).

A second design principle is the use of typed interfaces as the contract between code and model. Latent world state is represented through explicit types and schemas—JSON-like contracts or TypeScript interfaces—so that outputs can be validated, cached, rendered, and debugged. For open-ended scale, the framework uses deterministic generation from a stable seed derived from a hash of a location or identifier, with the stated persistence condition

$\omega(\ell,x)$7

This gives revisitable worlds without storing an infinite database (Feng et al., 29 Dec 2025).

The demonstrations cover several classes of web-native world: the Infinite Travel Atlas, grounded in real geography; the Galaxy Travel Atlas, with procedural galaxy structure and model-generated lore; Bookshelf, for long-form narrative continuity; AI Spire, where deterministic combat and deck logic coexist with typed, model-generated cards; AI Alchemy, which combines deterministic cellular automata with generated reaction rules; Cosmic Voyager, a WebGL planetary exploration world; and WWMPedia, where retrieval, sanitization, routing, and HTML rendering constitute the physics layer while the model writes a persistent encyclopedia-style artifact (Feng et al., 29 Dec 2025).

The broader implication is architectural. WebWorld, in this sense, is no longer only a simulator of browser transitions or a synthetic benchmark website. It becomes a disciplined way of building persistent interactive worlds on the web: code enforces state transitions and invariants, while LLMs provide contextual richness, narrative, or open-ended semantic expansion. This suggests that modern web stacks can function not merely as deployment targets for agents, but as the world-model substrate itself (Feng et al., 29 Dec 2025).